From 9a41ca6c93f2c7325928cbd804270c39e04289d8 Mon Sep 17 00:00:00 2001
From: Ivan <root@ivan.digital>
Date: Sun, 5 Jul 2026 08:06:26 +0200
Subject: [PATCH 1/2] Support Chatterbox multilingual frontends

---
 README.md                                     |   8 +-
 .../ChatterboxTTS/ChatterboxTTSModel.swift    |  17 +-
 Sources/ChatterboxTTS/MTLTokenizer.swift      | 217 +++++++++++++++++-
 Sources/OmniVoiceTTS/OmniVoiceTTSModel.swift  |   6 +-
 .../MTLTokenizerTests.swift                   |  96 +++++++-
 .../RoundtripE2ETests.swift                   |   8 +-
 .../OmniVoiceModelFacadeTests.swift           |   9 +-
 docs/models/chatterbox-tts.md                 |  21 +-
 8 files changed, 344 insertions(+), 38 deletions(-)

diff --git a/README.md b/README.md
index 5240f478..b88a33e2 100644
--- a/README.md
+++ b/README.md
@@ -46,8 +46,8 @@ On-device speech recognition, synthesis, and understanding for Mac and iOS. Runs
 - **[VibeVoice TTS](https://soniqo.audio/guides/vibevoice)** — Long-form / multi-speaker TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, up to 90-min podcast/audiobook synthesis, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/guides/magpie)** — Multilingual TTS (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB or CoreML INT8 342 MB, 9 languages, 5 baked speakers, streaming on MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — On-device flow-matching TTS (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 languages, 10 voices, G2P-free, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Multilingual TTS with zero-shot voice cloning (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 23 languages, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — Non-autoregressive diffusion TTS with zero-shot voice cloning (k2-fsa OmniVoice, Qwen3 backbone, MLX int8 ~1 GB / fp16, 600+ languages, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Multilingual TTS with zero-shot voice cloning (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 22 runtime languages, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — Non-autoregressive diffusion TTS with zero-shot voice cloning (k2-fsa OmniVoice, Qwen3 backbone, MLX fp16 default / int8 available, 600+ languages, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLMs & Translation**
@@ -168,8 +168,8 @@ Compact view below. **[Full model catalogue with sizes, quantisations, download
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/guides/vibevoice) | Text → Speech (long-form, multi-speaker) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/guides/vibevoice) | Text → Speech (up to 90-min podcast) | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/guides/magpie) | Text → Speech (5 baked speakers, streaming) | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9 (CoreML excludes JA) |
-| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Text → Speech (zero-shot cloning) | MLX | 0.8B (fp16) | 23 |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | Text → Speech (NAR diffusion, zero-shot cloning) | MLX | 0.8B (int8/fp16) | **600+** |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Text → Speech (zero-shot cloning) | MLX | 0.8B (fp16) | 22 runtime |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Text → Speech (NAR diffusion, zero-shot cloning) | MLX | 0.8B (fp16 default / int8) | **600+** |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Text → Speech (zero-shot cloning, explicit style markers) | MLX | 0.5B-class (fp16) | Multilingual |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/Sources/ChatterboxTTS/ChatterboxTTSModel.swift b/Sources/ChatterboxTTS/ChatterboxTTSModel.swift
index 9164e092..66508d6c 100644
--- a/Sources/ChatterboxTTS/ChatterboxTTSModel.swift
+++ b/Sources/ChatterboxTTS/ChatterboxTTSModel.swift
@@ -204,7 +204,16 @@ public final class ChatterboxTTSModel {
             for: s3TokenizerModelId, cacheDirName: "chatterbox-s3-tokenizer")
 
         let fm = FileManager.default
-        if !fm.fileExists(atPath: bundleDir.appendingPathComponent("model.safetensors").path) {
+        let requiredBundleFiles = [
+            "model.safetensors",
+            "config.json",
+            "tokenizer.json",
+            "Cangjie5_TC.json",
+        ]
+        let bundleNeedsRepair = requiredBundleFiles.contains {
+            !fm.fileExists(atPath: bundleDir.appendingPathComponent($0).path)
+        }
+        if bundleNeedsRepair {
             progressHandler?(0.0, "Downloading \(modelId)...")
             // No `.safetensors` in additionalFiles: that keeps downloadWeights'
             // automatic `*.safetensors` glob enabled, which fetches the bundle's
@@ -212,7 +221,7 @@ public final class ChatterboxTTSModel {
             // `.safetensors` here would disable that glob and drop model.safetensors.
             try await HuggingFaceDownloader.downloadWeights(
                 modelId: modelId, to: bundleDir,
-                additionalFiles: ["config.json", "tokenizer.json"],
+                additionalFiles: ["config.json", "tokenizer.json", "Cangjie5_TC.json"],
                 offlineMode: offlineMode
             ) { progressHandler?($0 * 0.7, "Downloading model...") }
         }
@@ -278,7 +287,7 @@ public final class ChatterboxTTSModel {
     ///   - referenceSamples: reference clip, mono.
     ///   - sampleRate: sample rate of `referenceSamples`.
     ///   - text: text to speak.
-    ///   - languageId: e.g. "en", "ar", "hi", "de", "es", "fr", "it", "pt".
+    ///   - languageId: one of ``MTLTokenizer/supportedLanguages``.
     ///   - exaggeration: emotion-advance scalar (T3 `emotion_adv`).
     ///   - cfgWeight: T3 classifier-free-guidance weight.
     ///   - temperature: T3 sampling temperature (0 = greedy).
@@ -297,7 +306,7 @@ public final class ChatterboxTTSModel {
         cfgWeight: Float = 0.5
     ) throws -> [Float] {
         let lang = languageId.lowercased()
-        guard MTLTokenizer.frontendFreeLanguages.contains(lang) else {
+        guard MTLTokenizer.supportedLanguages.contains(lang) else {
             throw ChatterboxModelError.unsupportedLanguage(lang)
         }
 
diff --git a/Sources/ChatterboxTTS/MTLTokenizer.swift b/Sources/ChatterboxTTS/MTLTokenizer.swift
index 35713e5f..5162a3ab 100644
--- a/Sources/ChatterboxTTS/MTLTokenizer.swift
+++ b/Sources/ChatterboxTTS/MTLTokenizer.swift
@@ -1,4 +1,6 @@
 import Foundation
+import CoreFoundation
+import NaturalLanguage
 
 /// Special tokens of the Chatterbox multilingual grapheme tokenizer.
 public enum ChatterboxToken {
@@ -10,12 +12,13 @@ public enum ChatterboxToken {
 
 public enum ChatterboxTokenizerError: Error {
     case malformedTokenizerJSON
+    case missingCangjieMap
 }
 
 /// Multilingual grapheme tokenizer for Chatterbox — a self-contained Swift port
-/// of the reference `MTLTokenizer`. The encode pipeline for Latin-script languages
-/// plus Arabic and Hindi is: lowercase → NFKD → prepend the `[lang]` token →
-/// replace spaces with `[SPACE]` → BPE.
+/// of the reference `MTLTokenizer`. The common encode pipeline is:
+/// lowercase → NFKD → prepend the `[lang]` token → replace spaces with
+/// `[SPACE]` → BPE.
 ///
 /// The BPE faithfully reproduces the HF `tokenizer.json` (model `BPE`,
 /// pre-tokenizer `Whitespace`, no normalizer, `[UNK]` fallback): added/control
@@ -25,20 +28,39 @@ public enum ChatterboxTokenizerError: Error {
 /// reference implementation exactly.
 ///
 /// Some writing systems additionally need a language-specific text frontend
-/// before this grapheme path — see `frontendFreeLanguages` for the set this
-/// tokenizer encodes directly.
+/// before this grapheme path. Runtime language support is intentionally limited
+/// to languages whose frontend is implemented here; unsupported upstream tags
+/// stay visible as metadata but are not accepted by synthesis.
 public final class MTLTokenizer {
     private let vocab: [String: Int]
     private let idToToken: [Int: String]
     private let mergeRank: [String: Int]
+    private let cangjie: [String: String]
+    private let cangjieDuplicateIndex: [String: Int]
     /// (content, id) of added/control tokens, sorted longest content first so the
     /// scan is greedy-longest-match like the HF tokenizer.
     private let addedTokens: [(token: [Character], id: Int)]
     private let unkId: Int
 
-    /// Languages this tokenizer encodes directly (no language-specific frontend).
-    public static let frontendFreeLanguages: Set<String> =
-        ["en", "ar", "hi", "de", "es", "fr", "it", "pt"]
+    /// Languages published by Resemble Chatterbox Multilingual V3.
+    public static let upstreamLanguages: Set<String> = [
+        "ar", "da", "de", "el", "en", "es", "fi", "fr", "he", "hi",
+        "it", "ja", "ko", "ms", "nl", "no", "pl", "pt", "ru", "sv",
+        "sw", "tr", "zh",
+    ]
+
+    /// Languages whose text frontend is implemented locally without falling
+    /// back to raw text. Keep languages with missing external frontends out of
+    /// this runtime gate: `he` needs Dicta-style diacritization.
+    public static let supportedLanguages: Set<String> = [
+        "ar", "da", "de", "el", "en", "es", "fi", "fr", "hi", "it",
+        "ja", "ko", "ms", "nl", "no", "pl", "pt", "ru", "sv", "sw",
+        "tr", "zh",
+    ]
+
+    /// Backwards-compatible alias retained for callers that used this as a
+    /// runtime language gate.
+    public static let frontendFreeLanguages: Set<String> = supportedLanguages
 
     private struct TokenizerJSON: Decodable {
         struct Model: Decodable {
@@ -69,6 +91,9 @@ public final class MTLTokenizer {
         var ranks = [String: Int](minimumCapacity: tj.model.merges.count)
         for (rank, merge) in tj.model.merges.enumerated() { ranks[merge] = rank }
         self.mergeRank = ranks
+        let cangjieMap = try Self.loadCangjieMap(from: modelFolder)
+        self.cangjie = cangjieMap.codes
+        self.cangjieDuplicateIndex = cangjieMap.duplicateIndex
 
         let unk = tj.model.unk_token ?? ChatterboxToken.unk
         self.unkId = tj.model.vocab[unk] ?? 1
@@ -82,11 +107,23 @@ public final class MTLTokenizer {
     /// Encode text to token ids. `languageId` (e.g. "en", "ar", "hi") is prepended
     /// as a `[lang]` control token, matching the multilingual model's training.
     public func encode(_ text: String, languageId: String? = nil) -> [Int] {
+        let lang = languageId?.trimmingCharacters(in: .whitespacesAndNewlines).lowercased()
         var s = text.lowercased()
-        // NFKD — Swift's compatibility decomposition matches Python `normalize("NFKD", …)`.
-        s = s.decomposedStringWithCompatibilityMapping
-        if let lang = languageId, !lang.isEmpty {
-            s = "[\(lang.lowercased())]" + s
+        if lang == "ja" {
+            s = Self.normalizeJapaneseToHiragana(s)
+            // Japanese reference normalization applies NFKD after hiragana conversion.
+            s = s.decomposedStringWithCompatibilityMapping
+        } else {
+            // NFKD — Swift's compatibility decomposition matches Python `normalize("NFKD", …)`.
+            s = s.decomposedStringWithCompatibilityMapping
+        }
+        if lang == "zh", !cangjie.isEmpty {
+            s = applyCangjieFrontend(Self.segmentChinese(s))
+        } else if lang == "ko" {
+            s = Self.decomposeHangulToJamo(s)
+        }
+        if let lang, !lang.isEmpty {
+            s = "[\(lang)]" + s
         }
         s = s.replacingOccurrences(of: " ", with: ChatterboxToken.space)
         return tokenize(s)
@@ -104,6 +141,162 @@ public final class MTLTokenizer {
 
     // MARK: - Internals
 
+    private static func loadCangjieMap(from modelFolder: URL) throws -> (
+        codes: [String: String],
+        duplicateIndex: [String: Int]
+    ) {
+        let url = modelFolder.appendingPathComponent("Cangjie5_TC.json")
+        guard let data = try? Data(contentsOf: url),
+              let rows = try? JSONDecoder().decode([String].self, from: data) else {
+            throw ChatterboxTokenizerError.missingCangjieMap
+        }
+        var codes: [String: String] = [:]
+        var glyphsByCode: [String: [String]] = [:]
+        codes.reserveCapacity(rows.count)
+        for row in rows {
+            let parts = row.split(separator: "\t", omittingEmptySubsequences: false)
+            guard parts.count >= 2 else { continue }
+            let glyph = String(parts[0])
+            let code = String(parts[1])
+            if codes[glyph] == nil {
+                codes[glyph] = code
+                glyphsByCode[code, default: []].append(glyph)
+            }
+        }
+        var duplicateIndex: [String: Int] = [:]
+        for glyphs in glyphsByCode.values {
+            for (index, glyph) in glyphs.enumerated() where index > 0 {
+                duplicateIndex[glyph] = index
+            }
+        }
+        return (codes, duplicateIndex)
+    }
+
+    private func applyCangjieFrontend(_ text: String) -> String {
+        var out = ""
+        out.reserveCapacity(text.count * 4)
+        for character in text.map(String.init) {
+            guard let code = cangjie[character] else {
+                out += character
+                continue
+            }
+            let indexedCode = if let index = cangjieDuplicateIndex[character] {
+                "\(code)\(index)"
+            } else {
+                code
+            }
+            for scalar in indexedCode.unicodeScalars {
+                out += "[cj_\(String(scalar))]"
+            }
+            out += "[cj_.]"
+        }
+        return out
+    }
+
+    private static func segmentChinese(_ text: String) -> String {
+        guard !text.isEmpty else { return text }
+        let tokenizer = NLTokenizer(unit: .word)
+        tokenizer.string = text
+        tokenizer.setLanguage(.simplifiedChinese)
+
+        var tokens: [String] = []
+        var cursor = text.startIndex
+
+        func appendGap(_ gap: Substring) {
+            for scalar in gap.unicodeScalars where !CharacterSet.whitespacesAndNewlines.contains(scalar) {
+                tokens.append(String(scalar))
+            }
+        }
+
+        tokenizer.enumerateTokens(in: text.startIndex..<text.endIndex) { range, _ in
+            if cursor < range.lowerBound {
+                appendGap(text[cursor..<range.lowerBound])
+            }
+            tokens.append(String(text[range]))
+            cursor = range.upperBound
+            return true
+        }
+        if cursor < text.endIndex {
+            appendGap(text[cursor..<text.endIndex])
+        }
+        return tokens.isEmpty ? text : tokens.joined(separator: " ")
+    }
+
+    private static func normalizeJapaneseToHiragana(_ text: String) -> String {
+        guard !text.isEmpty else { return text }
+        let cfText = text as CFString
+        let tokenizer = CFStringTokenizerCreate(
+            nil,
+            cfText,
+            CFRangeMake(0, CFStringGetLength(cfText)),
+            kCFStringTokenizerUnitWord,
+            Locale(identifier: "ja_JP") as CFLocale
+        )
+
+        var out = ""
+        var cursor = text.startIndex
+        var tokenType = CFStringTokenizerAdvanceToNextToken(tokenizer)
+        while tokenType != [] {
+            let cfRange = CFStringTokenizerGetCurrentTokenRange(tokenizer)
+            let nsRange = NSRange(location: cfRange.location, length: cfRange.length)
+            guard let swiftRange = Range(nsRange, in: text) else {
+                tokenType = CFStringTokenizerAdvanceToNextToken(tokenizer)
+                continue
+            }
+            if cursor < swiftRange.lowerBound {
+                out += text[cursor..<swiftRange.lowerBound]
+            }
+            let raw = String(text[swiftRange])
+            if Self.containsKanji(raw),
+               let latin = CFStringTokenizerCopyCurrentTokenAttribute(
+                    tokenizer,
+                    kCFStringTokenizerAttributeLatinTranscription
+               ) as? String,
+               let hiragana = latin.applyingTransform(.latinToHiragana, reverse: false),
+               !hiragana.isEmpty {
+                if let first = hiragana.first, first == "は" || first == "へ" {
+                    out += " "
+                }
+                out += hiragana
+            } else {
+                out += raw
+            }
+            cursor = swiftRange.upperBound
+            tokenType = CFStringTokenizerAdvanceToNextToken(tokenizer)
+        }
+        if cursor < text.endIndex {
+            out += text[cursor..<text.endIndex]
+        }
+        return out
+    }
+
+    private static func containsKanji(_ text: String) -> Bool {
+        text.unicodeScalars.contains { scalar in
+            (0x4E00...0x9FFF).contains(scalar.value)
+        }
+    }
+
+    private static func decomposeHangulToJamo(_ text: String) -> String {
+        var out = ""
+        for scalar in text.unicodeScalars {
+            let value = scalar.value
+            guard (0xAC00...0xD7A3).contains(value) else {
+                out.unicodeScalars.append(scalar)
+                continue
+            }
+            let base = value - 0xAC00
+            let initial = 0x1100 + base / (21 * 28)
+            let medial = 0x1161 + (base % (21 * 28)) / 28
+            let final = base % 28
+            out.unicodeScalars.append(UnicodeScalar(initial)!)
+            out.unicodeScalars.append(UnicodeScalar(medial)!)
+            if final > 0 {
+                out.unicodeScalars.append(UnicodeScalar(0x11A7 + final)!)
+            }
+        }
+        return out
+    }
+
     /// Split the preprocessed string by added/control tokens (longest match), then
     /// pre-tokenize each gap with the Whitespace rule and BPE each word.
     private func tokenize(_ s: String) -> [Int] {
diff --git a/Sources/OmniVoiceTTS/OmniVoiceTTSModel.swift b/Sources/OmniVoiceTTS/OmniVoiceTTSModel.swift
index 43b01a5c..a2e4eb49 100644
--- a/Sources/OmniVoiceTTS/OmniVoiceTTSModel.swift
+++ b/Sources/OmniVoiceTTS/OmniVoiceTTSModel.swift
@@ -8,9 +8,9 @@ import MLXNN
 /// Qwen3 diffusion backbone, the Higgs-audio v2 codec (encode + decode), and the
 /// text front-end. This is the entry point the Studio sidecar drives.
 public final class OmniVoiceTTSModel {
-    /// Default published bundle (int8 backbone + fp16 codec). The fp16 backbone
-    /// bundle is `aufklarer/OmniVoice-MLX-fp16`.
-    public static let defaultModelId = "aufklarer/OmniVoice-MLX-int8"
+    /// Default published bundle (fp16 backbone + fp16 codec). The int8 backbone
+    /// remains loadable by passing `aufklarer/OmniVoice-MLX-int8` explicitly.
+    public static let defaultModelId = OmniVoiceConfig.defaultModelId
 
     let cfg: OmniVoiceConfig
     let model: OmniVoiceModel
diff --git a/Tests/ChatterboxTTSTests/MTLTokenizerTests.swift b/Tests/ChatterboxTTSTests/MTLTokenizerTests.swift
index 4900803e..0aafaa02 100644
--- a/Tests/ChatterboxTTSTests/MTLTokenizerTests.swift
+++ b/Tests/ChatterboxTTSTests/MTLTokenizerTests.swift
@@ -4,13 +4,13 @@ import XCTest
 @testable import ChatterboxTTS
 
 /// Numeric gate for the Swift `MTLTokenizer`: token ids must match the
-/// reference golden exactly for the frontend-free languages we support first
-/// (en / ar / hi).
+/// reference golden exactly for core languages, and every runtime-supported
+/// language frontend must encode without unknown-token fallback.
 final class MTLTokenizerTests: XCTestCase {
     private func loadTokenizer() async throws -> MTLTokenizer {
         let folder = try await HubApi().snapshot(
             from: "aufklarer/Chatterbox-Multilingual-MLX-fp16",
-            matching: ["tokenizer.json"]
+            matching: ["tokenizer.json", "Cangjie5_TC.json"]
         )
         return try MTLTokenizer(modelFolder: folder)
     }
@@ -37,4 +37,94 @@ final class MTLTokenizerTests: XCTestCase {
         XCTAssertEqual(tok.encode("a", languageId: "ar").first, 721)
         XCTAssertEqual(tok.encode("a", languageId: "hi").first, 722)
     }
+
+    func testLanguageSetsSeparateUpstreamFromRuntimeReadyFrontends() async throws {
+        let expected: Set<String> = [
+            "ar", "da", "de", "el", "en", "es", "fi", "fr", "he", "hi",
+            "it", "ja", "ko", "ms", "nl", "no", "pl", "pt", "ru", "sv",
+            "sw", "tr", "zh",
+        ]
+        XCTAssertEqual(MTLTokenizer.upstreamLanguages, expected)
+        XCTAssertEqual(MTLTokenizer.upstreamLanguages.count, 23)
+        XCTAssertFalse(MTLTokenizer.supportedLanguages.contains("he"))
+        XCTAssertTrue(MTLTokenizer.supportedLanguages.contains("ja"))
+        XCTAssertTrue(MTLTokenizer.supportedLanguages.contains("zh"))
+
+        let tok = try await loadTokenizer()
+        for language in MTLTokenizer.supportedLanguages.sorted() {
+            let encoded = tok.encode("hello", languageId: language)
+            XCTAssertFalse(encoded.isEmpty, "language \(language) should encode")
+            XCTAssertEqual(
+                tok.decode([encoded[0]]),
+                "[\(language)]",
+                "language \(language) should prepend its control token")
+        }
+    }
+
+    func testChineseUsesCangjieFrontendWhenMapIsAvailable() async throws {
+        let tok = try await loadTokenizer()
+        let encoded = tok.encode("今天", languageId: "zh")
+
+        XCTAssertEqual(encoded.first, 725)
+        XCTAssertEqual(
+            tok.decode(Array(encoded.dropFirst())),
+            "[cj_o][cj_i][cj_n][cj_.][cj_m][cj_k][cj_.]")
+        XCTAssertFalse(encoded.contains(1), "Cangjie-mapped Chinese should not fall back to [UNK]")
+    }
+
+    func testChineseFrontendSegmentsWordsBeforeCangjie() async throws {
+        let tok = try await loadTokenizer()
+        let encoded = tok.encode("今天我们测试中文语音", languageId: "zh")
+        let decoded = tok.decode(Array(encoded.dropFirst()))
+
+        XCTAssertTrue(decoded.contains(" "), "Chinese frontend should preserve word boundaries")
+        XCTAssertTrue(decoded.contains("[cj_"), "Chinese text should be converted to Cangjie tokens")
+        XCTAssertFalse(encoded.contains(1), "Segmented Chinese should not fall back to [UNK]")
+    }
+
+    func testJapaneseUsesHiraganaFrontendForKanji() async throws {
+        let tok = try await loadTokenizer()
+        let encoded = tok.encode("今日はカタカナです。", languageId: "ja")
+
+        XCTAssertEqual(encoded.first, 723)
+        XCTAssertTrue(
+            tok.decode(Array(encoded.dropFirst())).contains("きょうはカタカナです"),
+            "Kanji should be normalized to hiragana while katakana stays unchanged"
+        )
+        XCTAssertFalse(encoded.contains(1), "Japanese frontend should not fall back to [UNK]")
+    }
+
+    func testBenchmarkSentencesDoNotEmitUnknownTokens() async throws {
+        let tok = try await loadTokenizer()
+        let samples = [
+            "ar": "اليوم نختبر ما إذا كان الكلام واضحا وسهل الفهم.",
+            "da": "I dag tester vi, om talen er klar og let at forstå.",
+            "de": "Heute testen wir, ob die Sprache klar und leicht zu verstehen ist.",
+            "el": "Σήμερα ελέγχουμε αν η ομιλία είναι καθαρή και εύκολη στην κατανόηση.",
+            "en": "Today we test whether the speech is clear and easy to understand.",
+            "es": "Hoy probamos si el habla es clara y fácil de entender.",
+            "fi": "Tänään testaamme, onko puhe selkeää ja helppoa ymmärtää.",
+            "fr": "Aujourd'hui, nous testons si la parole est claire et facile à comprendre.",
+            "hi": "आज हम जाँचते हैं कि आवाज़ साफ़ और समझने में आसान है या नहीं।",
+            "it": "Oggi testiamo se il parlato è chiaro e facile da capire.",
+            "ja": "今日は音声が明瞭で理解しやすいかを確認します。",
+            "ko": "오늘 우리는 음성이 명확하고 이해하기 쉬운지 테스트합니다.",
+            "ms": "Hari ini kami menguji sama ada pertuturan jelas dan mudah difahami.",
+            "nl": "Vandaag testen we of de spraak duidelijk en gemakkelijk te begrijpen is.",
+            "no": "I dag tester vi om talen er klar og lett å forstå.",
+            "pl": "Dzisiaj sprawdzamy, czy mowa jest wyraźna i łatwa do zrozumienia.",
+            "pt": "Hoje testamos se a fala é clara e fácil de entender.",
+            "ru": "Сегодня мы проверяем, является ли речь четкой и легкой для понимания.",
+            "sv": "I dag testar vi om talet är tydligt och lätt att förstå.",
+            "sw": "Leo tunajaribu kama hotuba iko wazi na ni rahisi kuelewa.",
+            "tr": "Bugün konuşmanın net ve kolay anlaşılır olup olmadığını test ediyoruz.",
+            "zh": "今天我们测试中文语音是否清楚且容易理解。",
+        ]
+
+        XCTAssertEqual(Set(samples.keys), MTLTokenizer.supportedLanguages)
+        for (language, text) in samples {
+            let encoded = tok.encode(text, languageId: language)
+            XCTAssertFalse(encoded.contains(1), "\(language) should not emit [UNK]")
+        }
+    }
 }
diff --git a/Tests/ChatterboxTTSTests/RoundtripE2ETests.swift b/Tests/ChatterboxTTSTests/RoundtripE2ETests.swift
index 3498b99f..4a0fd6d2 100644
--- a/Tests/ChatterboxTTSTests/RoundtripE2ETests.swift
+++ b/Tests/ChatterboxTTSTests/RoundtripE2ETests.swift
@@ -11,7 +11,7 @@ import XCTest
 /// and the reference speaker embedding — is buildable today and gated here.
 /// The full text→audio→verify roundtrip is scaffolded and activates once the T3
 /// (text→speech-token) and S3Gen (speech-token→waveform) stages land.
-final class RoundtripE2ETests: XCTestCase {
+final class E2ERoundtripTests: XCTestCase {
     private let veWeights = "/tmp/cbx_ve.safetensors"
     private let veInput = "/tmp/cbx_ve_input16k.f32"
 
@@ -79,8 +79,8 @@ final class RoundtripE2ETests: XCTestCase {
                           to: "/tmp/cbx_swift_synth.wav")
     }
 
-    /// Synthesize across languages (en/ar/hi — the multilingual point of Chatterbox)
-    /// and write each to /tmp/cbx_swift_<lang>.wav for ASR verification.
+    /// Synthesize across languages, including frontend-heavy scripts, and write
+    /// each to /tmp/cbx_swift_<lang>.wav for ASR verification.
     func testMultilingualSynthesis() throws {
         let fm = FileManager.default
         let bundleDir = "/tmp/cbx-fp16"
@@ -101,6 +101,8 @@ final class RoundtripE2ETests: XCTestCase {
             ("en", "Hello there, this is a cloned voice."),
             ("ar", "مرحبا، هذا صوت مستنسخ."),
             ("hi", "नमस्ते, यह एक क्लोन की गई आवाज़ है।"),
+            ("ja", "今日は音声が明瞭で理解しやすいかを確認します。"),
+            ("zh", "今天我们测试中文语音是否清楚且容易理解。"),
         ]
         for (lang, text) in cases {
             let audio = try model.clone(
diff --git a/Tests/OmniVoiceTTSTests/OmniVoiceModelFacadeTests.swift b/Tests/OmniVoiceTTSTests/OmniVoiceModelFacadeTests.swift
index 064d107a..58d0452b 100644
--- a/Tests/OmniVoiceTTSTests/OmniVoiceModelFacadeTests.swift
+++ b/Tests/OmniVoiceTTSTests/OmniVoiceModelFacadeTests.swift
@@ -6,9 +6,14 @@ import XCTest
 
 /// Exercises the high-level `OmniVoiceTTSModel` facade the Studio sidecar drives:
 /// load a local bundle, then synthesize cloned speech from a reference WAV + text.
-/// Writes the waveform for an external ASR check. Uses the int8 bundle so the
-/// quantization-from-config load path is covered too.
+/// Writes the waveform for an external ASR check. The optional bundle test below
+/// still uses int8 so the quantization-from-config load path remains covered.
 final class OmniVoiceModelFacadeTests: XCTestCase {
+    func testDefaultModelIdUsesFp16Bundle() {
+        XCTAssertEqual(OmniVoiceTTSModel.defaultModelId, OmniVoiceConfig.defaultModelId)
+        XCTAssertEqual(OmniVoiceTTSModel.defaultModelId, "aufklarer/OmniVoice-MLX-fp16")
+    }
+
     func testFromBundleAndGenerate() async throws {
         let bundleDir = "/tmp/omnivoice-int8"
         let refWav = "/tmp/cbx_swift_en.wav"
diff --git a/docs/models/chatterbox-tts.md b/docs/models/chatterbox-tts.md
index 8a6519fc..dcac3e69 100644
--- a/docs/models/chatterbox-tts.md
+++ b/docs/models/chatterbox-tts.md
@@ -3,9 +3,10 @@
 ## Overview
 
 Chatterbox Multilingual (Resemble AI, MIT) is a zero-shot voice-cloning TTS that
-synthesizes a target text in the timbre of a short reference clip across 23
-languages — the broadest language coverage of any TTS in this package. The MLX
-bundle is a genuine fp16
+synthesizes a target text in the timbre of a short reference clip. The upstream
+model publishes 23 language tags; the Swift runtime enables the 22 languages
+whose required text frontends are implemented locally. The MLX bundle is a
+genuine fp16
 conversion (~1.3 GB) of the three upstream checkpoints, published at
 [`aufklarer/Chatterbox-Multilingual-MLX-fp16`](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16).
 
@@ -14,12 +15,18 @@ speaker embeddings checked against a known-good reference at each stage.
 
 ## Languages
 
-23 languages: Arabic, Chinese, Danish, Dutch, English, Finnish, French, German,
-Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish,
-Portuguese, Russian, Spanish, Swahili, Swedish, Turkish.
+Runtime-enabled languages: Arabic, Chinese, Danish, Dutch, English, Finnish,
+French, German, Greek, Hindi, Italian, Japanese, Korean, Malay, Norwegian,
+Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Turkish.
 
 Text encodes through a grapheme path: NFKD normalization → a per-language
-`[lang]` token → BPE.
+`[lang]` token → BPE. Frontend-heavy scripts are handled before that path:
+
+- Chinese uses word segmentation plus the published Cangjie-5 token map.
+- Japanese converts kanji readings to hiragana through `CFStringTokenizer`.
+- Korean decomposes Hangul syllables to Jamo.
+- Hebrew remains disabled until the runtime includes the diacritization frontend
+  needed by the upstream model.
 
 ## Pipeline Stages
 

From 1c31dd27f993a36f66827964993da6e7ee522a96 Mon Sep 17 00:00:00 2001
From: Ivan <root@ivan.digital>
Date: Sun, 5 Jul 2026 08:39:31 +0200
Subject: [PATCH 2/2] Add safe Hebrew support for Chatterbox

---
 README.md                                     |  4 +-
 README_ar.md                                  |  8 +--
 README_de.md                                  |  8 +--
 README_es.md                                  |  8 +--
 README_fr.md                                  |  8 +--
 README_hi.md                                  |  8 +--
 README_ja.md                                  |  7 ++-
 README_ko.md                                  |  7 ++-
 README_pt.md                                  |  7 ++-
 README_ru.md                                  |  7 ++-
 README_th.md                                  |  7 ++-
 README_tr.md                                  |  7 ++-
 README_vi.md                                  |  7 ++-
 README_zh.md                                  |  7 ++-
 .../ChatterboxTTS/ChatterboxTTSModel.swift    | 11 +++-
 Sources/ChatterboxTTS/MTLTokenizer.swift      | 56 +++++++++++++++++--
 .../MTLTokenizerTests.swift                   | 25 ++++++++-
 .../RoundtripE2ETests.swift                   |  1 +
 docs/models/chatterbox-tts.md                 | 14 ++---
 19 files changed, 146 insertions(+), 61 deletions(-)

diff --git a/README.md b/README.md
index b88a33e2..59139b35 100644
--- a/README.md
+++ b/README.md
@@ -46,7 +46,7 @@ On-device speech recognition, synthesis, and understanding for Mac and iOS. Runs
 - **[VibeVoice TTS](https://soniqo.audio/guides/vibevoice)** — Long-form / multi-speaker TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, up to 90-min podcast/audiobook synthesis, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/guides/magpie)** — Multilingual TTS (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB or CoreML INT8 342 MB, 9 languages, 5 baked speakers, streaming on MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — On-device flow-matching TTS (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 languages, 10 voices, G2P-free, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Multilingual TTS with zero-shot voice cloning (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 22 runtime languages, MIT)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Multilingual TTS with zero-shot voice cloning (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 23 runtime languages; Hebrew requires niqqud, MIT)
 - **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — Non-autoregressive diffusion TTS with zero-shot voice cloning (k2-fsa OmniVoice, Qwen3 backbone, MLX fp16 default / int8 available, 600+ languages, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
@@ -168,7 +168,7 @@ Compact view below. **[Full model catalogue with sizes, quantisations, download
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/guides/vibevoice) | Text → Speech (long-form, multi-speaker) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/guides/vibevoice) | Text → Speech (up to 90-min podcast) | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/guides/magpie) | Text → Speech (5 baked speakers, streaming) | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9 (CoreML excludes JA) |
-| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Text → Speech (zero-shot cloning) | MLX | 0.8B (fp16) | 22 runtime |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Text → Speech (zero-shot cloning) | MLX | 0.8B (fp16) | 23 (HE requires niqqud) |
 | [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Text → Speech (NAR diffusion, zero-shot cloning) | MLX | 0.8B (fp16 default / int8) | **600+** |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Text → Speech (zero-shot cloning, explicit style markers) | MLX | 0.5B-class (fp16) | Multilingual |
diff --git a/README_ar.md b/README_ar.md
index e3dc4883..4f4c1527 100644
--- a/README_ar.md
+++ b/README_ar.md
@@ -56,8 +56,8 @@
 - **[VibeVoice TTS](https://soniqo.audio/ar/guides/vibevoice)** — تحويل النص إلى كلام للنصوص الطويلة / متعدد المتحدثين (Microsoft VibeVoice Realtime-0.5B + 1.5B، MLX، توليد بودكاست/كتب صوتية حتى 90 دقيقة، EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/ar/guides/magpie)** — تحويل النص إلى كلام متعدد اللغات (NVIDIA Magpie-TTS Multilingual 357M، MLX INT8 411 ميغابايت أو CoreML INT8 342 ميغابايت، 9 لغات، 5 متحدثين جاهزين، تدفق على MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — تحويل النص إلى كلام على الجهاز بمطابقة التدفق (Supertone Supertonic-3 99M، CoreML/Neural Engine، 31 لغة، 10 أصوات، بدون G2P، 44.1 كيلوهرتز)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — تحويل النص إلى كلام متعدد اللغات مع استنساخ الصوت بدون تدريب مسبق (Resemble AI Chatterbox Multilingual، MLX fp16 ~1.3 غيغابايت، 23 لغة، MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — تحويل النص إلى كلام بالانتشار غير الانحداري التلقائي مع استنساخ الصوت بدون تدريب مسبق (k2-fsa OmniVoice، عمود Qwen3 الفقري، MLX int8 ~1 GB / fp16، أكثر من 600 لغة، Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — تحويل النص إلى كلام متعدد اللغات مع استنساخ الصوت بدون تدريب مسبق (Resemble AI Chatterbox Multilingual، MLX fp16 ~1.3 غيغابايت، 23 لغة وقت تشغيل؛ العبرية تتطلب النِّقُود، MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — تحويل النص إلى كلام بالانتشار غير الانحداري التلقائي مع استنساخ الصوت بدون تدريب مسبق (k2-fsa OmniVoice، عمود Qwen3 الفقري، MLX fp16 افتراضي / int8 متاح، أكثر من 600 لغة، Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **نماذج LLM والترجمة**
@@ -207,8 +207,8 @@ struct DictateView: View {
 | [Qwen3-TTS](https://soniqo.audio/ar/guides/speak) | نص → كلام | MLX, CoreML | 0.6B, 1.7B | 10 |
 | [CosyVoice3](https://soniqo.audio/ar/guides/cosyvoice) | نص → كلام | MLX | 0.5B | 9 |
 | [VoxCPM2](https://soniqo.audio/ar/speech-generation) | نص → كلام (48 كيلوهرتز، تصميم + استنساخ صوت) | MLX | 2B (bf16/int8) | 30 |
-| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | نص → كلام (استنساخ صوت بدون تدريب) | MLX | 0.8B (fp16) | 23 |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | نص → كلام (انتشار NAR، استنساخ بدون تدريب) | MLX | 0.8B (int8/fp16) | 600+ |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | نص → كلام (استنساخ صوت بدون تدريب) | MLX | 0.8B (fp16) | 23 (HE تتطلب النِّقُود) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | نص → كلام (انتشار NAR، استنساخ بدون تدريب) | MLX | 0.8B (fp16 افتراضي / int8) | 600+ |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | نص → كلام (استنساخ صوت بدون تدريب، علامات نمط صريحة) | MLX | 0.5B-class (fp16) | متعدد اللغات |
 | [Kokoro-82M](https://soniqo.audio/ar/guides/kokoro) | نص → كلام | CoreML (ANE) | 82M | 10 |
diff --git a/README_de.md b/README_de.md
index 0a3c126f..856b77b7 100644
--- a/README_de.md
+++ b/README_de.md
@@ -46,8 +46,8 @@ Spracherkennung, -synthese und -verständnis auf dem Gerät für Mac und iOS. L
 - **[VibeVoice TTS](https://soniqo.audio/de/guides/vibevoice)** — Langform-/Multi-Speaker-TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, bis zu 90 Min. Podcast-/Hörbuch-Synthese, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/de/guides/magpie)** — Mehrsprachiges TTS (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB oder CoreML INT8 342 MB, 9 Sprachen, 5 vordefinierte Sprecher, Streaming auf MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — Flow-Matching-TTS auf dem Gerät (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 Sprachen, 10 Stimmen, G2P-frei, 44,1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Mehrsprachiges TTS mit Zero-Shot-Stimmklonen (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 Sprachen, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — Nicht-autoregressives Diffusions-TTS mit Zero-Shot-Stimmklonen (k2-fsa OmniVoice, Qwen3-Backbone, MLX int8 ~1 GB / fp16, 600+ Sprachen, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Mehrsprachiges TTS mit Zero-Shot-Stimmklonen (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 Laufzeitsprachen; Hebräisch erfordert Nikkud, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — Nicht-autoregressives Diffusions-TTS mit Zero-Shot-Stimmklonen (k2-fsa OmniVoice, Qwen3-Backbone, MLX fp16 Standard / int8 verfügbar, 600+ Sprachen, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLMs und Übersetzung**
@@ -163,8 +163,8 @@ Kompakte Übersicht unten. **[Vollständiger Modellkatalog mit Größen, Quantis
 | [Qwen3-TTS](https://soniqo.audio/de/guides/speak) | Text → Sprache | MLX, CoreML | 0.6B, 1.7B | 10 |
 | [CosyVoice3](https://soniqo.audio/de/guides/cosyvoice) | Text → Sprache | MLX | 0.5B | 9 |
 | [VoxCPM2](https://soniqo.audio/de/speech-generation) | Text → Sprache (48 kHz, Voice Design + Klonen) | MLX | 2B (bf16/int8) | 30 |
-| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Text → Sprache (Zero-Shot-Klonen) | MLX | 0.8B (fp16) | 23 |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | Text → Sprache (NAR-Diffusion, Zero-Shot-Klonen) | MLX | 0.8B (int8/fp16) | 600+ |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Text → Sprache (Zero-Shot-Klonen) | MLX | 0.8B (fp16) | 23 (HE erfordert Nikkud) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Text → Sprache (NAR-Diffusion, Zero-Shot-Klonen) | MLX | 0.8B (fp16 Standard / int8) | 600+ |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Text → Sprache (Zero-Shot-Klonen, explizite Stilmarker) | MLX | 0.5B-class (fp16) | Mehrsprachig |
 | [Kokoro-82M](https://soniqo.audio/de/guides/kokoro) | Text → Sprache | CoreML (ANE) | 82M | 10 |
diff --git a/README_es.md b/README_es.md
index 805322b4..1ab39769 100644
--- a/README_es.md
+++ b/README_es.md
@@ -46,8 +46,8 @@ Reconocimiento, síntesis y comprensión de voz en el dispositivo para Mac e iOS
 - **[VibeVoice TTS](https://soniqo.audio/es/guides/vibevoice)** — TTS de formato largo / múltiples hablantes (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, síntesis de podcast/audiolibro de hasta 90 min, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/es/guides/magpie)** — TTS multilingüe (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB o CoreML INT8 342 MB, 9 idiomas, 5 hablantes preconfigurados, streaming en MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — TTS de flow-matching en el dispositivo (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 idiomas, 10 voces, sin G2P, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS multilingüe con clonación de voz zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 idiomas, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — TTS de difusión no autorregresivo con clonación de voz zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX int8 ~1 GB / fp16, 600+ idiomas, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS multilingüe con clonación de voz zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 idiomas en runtime; hebreo requiere niqqud, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — TTS de difusión no autorregresivo con clonación de voz zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX fp16 por defecto / int8 disponible, 600+ idiomas, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM y traducción**
@@ -163,8 +163,8 @@ Vista compacta a continuación. **[Catálogo completo de modelos con tamaños, c
 | [Qwen3-TTS](https://soniqo.audio/es/guides/speak) | Texto → Voz | MLX, CoreML | 0.6B, 1.7B | 10 |
 | [CosyVoice3](https://soniqo.audio/es/guides/cosyvoice) | Texto → Voz | MLX | 0.5B | 9 |
 | [VoxCPM2](https://soniqo.audio/es/speech-generation) | Texto → Voz (48 kHz, diseño de voz + clonación) | MLX | 2B (bf16/int8) | 30 |
-| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Texto → Voz (clonación zero-shot) | MLX | 0.8B (fp16) | 23 |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | Texto → Voz (difusión NAR, clonación zero-shot) | MLX | 0.8B (int8/fp16) | 600+ |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Texto → Voz (clonación zero-shot) | MLX | 0.8B (fp16) | 23 (HE requiere niqqud) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Texto → Voz (difusión NAR, clonación zero-shot) | MLX | 0.8B (fp16 por defecto / int8) | 600+ |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Texto → Voz (clonación zero-shot, marcadores de estilo explícitos) | MLX | 0.5B-class (fp16) | Multilingüe |
 | [Kokoro-82M](https://soniqo.audio/es/guides/kokoro) | Texto → Voz | CoreML (ANE) | 82M | 10 |
diff --git a/README_fr.md b/README_fr.md
index 92dab454..84bdafad 100644
--- a/README_fr.md
+++ b/README_fr.md
@@ -46,8 +46,8 @@ Reconnaissance, synthese et comprehension vocale embarquees pour Mac et iOS. S'e
 - **[VibeVoice TTS](https://soniqo.audio/fr/guides/vibevoice)** -- TTS long format / multi-locuteurs (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, synthese de podcast/livre audio jusqu'a 90 min, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/fr/guides/magpie)** — TTS multilingue (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 Mo ou CoreML INT8 342 Mo, 9 langues, 5 voix prédéfinies, streaming sur MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — TTS embarque par appariement de flux (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 langues, 10 voix, sans G2P, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS multilingue avec clonage vocal zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 Go, 23 langues, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — TTS à diffusion non autorégressive avec clonage vocal zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX int8 ~1 Go / fp16, 600+ langues, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS multilingue avec clonage vocal zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 Go, 23 langues runtime; l'hébreu exige le niqqud, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — TTS à diffusion non autorégressive avec clonage vocal zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX fp16 par défaut / int8 disponible, 600+ langues, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM et traduction**
@@ -163,8 +163,8 @@ Vue compacte ci-dessous. **[Catalogue complet des modeles avec tailles, quantifi
 | [Qwen3-TTS](https://soniqo.audio/fr/guides/speak) | Texte → Parole | MLX, CoreML | 0.6B, 1.7B | 10 |
 | [CosyVoice3](https://soniqo.audio/fr/guides/cosyvoice) | Texte → Parole | MLX | 0.5B | 9 |
 | [VoxCPM2](https://soniqo.audio/fr/speech-generation) | Texte → Parole (48 kHz, conception vocale + clonage) | MLX | 2B (bf16/int8) | 30 |
-| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Texte → Voix (clonage zero-shot) | MLX | 0.8B (fp16) | 23 |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | Texte → Voix (diffusion NAR, clonage zero-shot) | MLX | 0.8B (int8/fp16) | 600+ |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Texte → Voix (clonage zero-shot) | MLX | 0.8B (fp16) | 23 (HE exige le niqqud) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Texte → Voix (diffusion NAR, clonage zero-shot) | MLX | 0.8B (fp16 par défaut / int8) | 600+ |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Texte → Voix (clonage zero-shot, marqueurs de style explicites) | MLX | 0.5B-class (fp16) | Multilingue |
 | [Kokoro-82M](https://soniqo.audio/fr/guides/kokoro) | Texte → Parole | CoreML (ANE) | 82M | 10 |
diff --git a/README_hi.md b/README_hi.md
index d3beedc8..30828bd0 100644
--- a/README_hi.md
+++ b/README_hi.md
@@ -46,8 +46,8 @@ Mac और iOS के लिए ऑन-डिवाइस स्पीच रि
 - **[VibeVoice TTS](https://soniqo.audio/hi/guides/vibevoice)** — लंबे-रूप / बहु-वक्ता TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, 90 मिनट तक के पॉडकास्ट / ऑडियोबुक संश्लेषण, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/hi/guides/magpie)** — बहुभाषी TTS (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB या CoreML INT8 342 MB, 9 भाषाएँ, 5 पूर्व-निर्धारित वक्ता, MLX पर स्ट्रीमिंग)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — ऑन-डिवाइस फ़्लो-मैचिंग TTS (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 भाषाएँ, 10 वॉयस, G2P-free, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — ज़ीरो-शॉट वॉयस क्लोनिंग के साथ बहुभाषी TTS (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 23 भाषाएँ, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — ज़ीरो-शॉट वॉयस क्लोनिंग के साथ नॉन-ऑटोरिग्रेसिव डिफ्यूज़न TTS (k2-fsa OmniVoice, Qwen3 बैकबोन, MLX int8 ~1 GB / fp16, 600+ भाषाएँ, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — ज़ीरो-शॉट वॉयस क्लोनिंग के साथ बहुभाषी TTS (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 23 रनटाइम भाषाएँ; हिब्रू के लिए niqqud आवश्यक, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — ज़ीरो-शॉट वॉयस क्लोनिंग के साथ नॉन-ऑटोरिग्रेसिव डिफ्यूज़न TTS (k2-fsa OmniVoice, Qwen3 बैकबोन, MLX fp16 डिफ़ॉल्ट / int8 उपलब्ध, 600+ भाषाएँ, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM और अनुवाद**
@@ -163,8 +163,8 @@ struct DictateView: View {
 | [Qwen3-TTS](https://soniqo.audio/hi/guides/speak) | टेक्स्ट → स्पीच | MLX, CoreML | 0.6B, 1.7B | 10 |
 | [CosyVoice3](https://soniqo.audio/hi/guides/cosyvoice) | टेक्स्ट → स्पीच | MLX | 0.5B | 9 |
 | [VoxCPM2](https://soniqo.audio/hi/speech-generation) | टेक्स्ट → स्पीच (48 kHz, वॉयस डिज़ाइन + क्लोनिंग) | MLX | 2B (bf16/int8) | 30 |
-| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | टेक्स्ट → स्पीच (ज़ीरो-शॉट क्लोनिंग) | MLX | 0.8B (fp16) | 23 |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | टेक्स्ट → स्पीच (NAR डिफ्यूज़न, ज़ीरो-शॉट क्लोनिंग) | MLX | 0.8B (int8/fp16) | 600+ |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | टेक्स्ट → स्पीच (ज़ीरो-शॉट क्लोनिंग) | MLX | 0.8B (fp16) | 23 (HE के लिए niqqud आवश्यक) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | टेक्स्ट → स्पीच (NAR डिफ्यूज़न, ज़ीरो-शॉट क्लोनिंग) | MLX | 0.8B (fp16 डिफ़ॉल्ट / int8) | 600+ |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | टेक्स्ट → स्पीच (ज़ीरो-शॉट क्लोनिंग, स्पष्ट स्टाइल मार्कर) | MLX | 0.5B-class (fp16) | बहुभाषी |
 | [Kokoro-82M](https://soniqo.audio/hi/guides/kokoro) | टेक्स्ट → स्पीच | CoreML (ANE) | 82M | 10 |
diff --git a/README_ja.md b/README_ja.md
index 9022cee4..5d19a360 100644
--- a/README_ja.md
+++ b/README_ja.md
@@ -46,8 +46,8 @@ Mac・iOS向けのオンデバイス音声認識・合成・理解。Apple Silic
 - **[VibeVoice TTS](https://soniqo.audio/ja/guides/vibevoice)** — 長尺・マルチスピーカーTTS（Microsoft VibeVoice Realtime-0.5B + 1.5B、MLX、最長90分のポッドキャスト／オーディオブック生成、EN/ZH）
 - **[Magpie TTS](https://soniqo.audio/ja/guides/magpie)** — 多言語 TTS（NVIDIA Magpie-TTS Multilingual 357M、MLX INT8 411 MB または CoreML INT8 342 MB、9 言語、5 つの組み込みスピーカー、MLX でストリーミング）
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — オンデバイスのフローマッチングTTS（Supertone Supertonic-3 99M、CoreML/Neural Engine、31言語、10ボイス、G2P-free、44.1 kHz）
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — ゼロショット音声クローン対応の多言語TTS（Resemble AI Chatterbox Multilingual、MLX fp16 ~1.3 GB、23言語、MIT）
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — ゼロショット音声クローンに対応した非自己回帰拡散TTS（k2-fsa OmniVoice、Qwen3バックボーン、MLX int8 ~1 GB / fp16、600+言語、Apache-2.0）
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — ゼロショット音声クローン対応の多言語TTS（Resemble AI Chatterbox Multilingual、MLX fp16 ~1.3 GB、23ランタイム言語、ヘブライ語はニクダーが必要、MIT）
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — ゼロショット音声クローンに対応した非自己回帰拡散TTS（k2-fsa OmniVoice、Qwen3バックボーン、MLX fp16デフォルト / int8利用可、600+言語、Apache-2.0）
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM と翻訳**
@@ -168,7 +168,8 @@ struct DictateView: View {
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/ja/guides/vibevoice) | テキスト → 音声（長尺・マルチスピーカー） | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/ja/guides/vibevoice) | テキスト → 音声（最長90分のポッドキャスト） | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/ja/guides/magpie) | テキスト → 音声（5 つの組み込みスピーカー、ストリーミング） | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9（CoreML は日本語を除く） |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | テキスト → 音声（NAR 拡散、ゼロショットクローン） | MLX | 0.8B (int8/fp16) | **600+** |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | テキスト → 音声（ゼロショットクローン） | MLX | 0.8B (fp16) | 23（HEはニクダー必須） |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | テキスト → 音声（NAR 拡散、ゼロショットクローン） | MLX | 0.8B (fp16デフォルト / int8) | **600+** |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | テキスト → 音声（ゼロショットクローン、明示的スタイルマーカー） | MLX | 0.5B-class (fp16) | 多言語 |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/README_ko.md b/README_ko.md
index 125551a0..7a80838a 100644
--- a/README_ko.md
+++ b/README_ko.md
@@ -46,8 +46,8 @@ Mac과 iOS를 위한 온디바이스 음성 인식, 합성 및 이해. Apple Sil
 - **[VibeVoice TTS](https://soniqo.audio/ko/guides/vibevoice)** — 장문 / 멀티 스피커 TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, 최대 90분 팟캐스트 / 오디오북 합성, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/ko/guides/magpie)** — 다국어 TTS (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB 또는 CoreML INT8 342 MB, 9개 언어, 5개 내장 스피커, MLX 스트리밍)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — 온디바이스 flow-matching TTS (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31개 언어, 10개 음색, G2P-free, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — 제로샷 음성 복제를 지원하는 다국어 TTS (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 23개 언어, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — 제로샷 음성 복제를 지원하는 비자기회귀 확산 TTS (k2-fsa OmniVoice, Qwen3 백본, MLX int8 ~1 GB / fp16, 600+개 언어, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — 제로샷 음성 복제를 지원하는 다국어 TTS (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 런타임 23개 언어, 히브리어는 niqqud 필요, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — 제로샷 음성 복제를 지원하는 비자기회귀 확산 TTS (k2-fsa OmniVoice, Qwen3 백본, MLX fp16 기본 / int8 사용 가능, 600+개 언어, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM 및 번역**
@@ -168,7 +168,8 @@ struct DictateView: View {
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/ko/guides/vibevoice) | 텍스트 → 음성 (장문, 멀티 스피커) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/ko/guides/vibevoice) | 텍스트 → 음성 (최대 90분 팟캐스트) | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/ko/guides/magpie) | 텍스트 → 음성 (5개 내장 스피커, 스트리밍) | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9 (CoreML은 일본어 제외) |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | 텍스트 → 음성 (NAR 확산, 제로샷 복제) | MLX | 0.8B (int8/fp16) | **600+** |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | 텍스트 → 음성 (제로샷 복제) | MLX | 0.8B (fp16) | 23 (HE는 niqqud 필요) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | 텍스트 → 음성 (NAR 확산, 제로샷 복제) | MLX | 0.8B (fp16 기본 / int8) | **600+** |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | 텍스트 → 음성 (제로샷 복제, 명시적 스타일 마커) | MLX | 0.5B-class (fp16) | 다국어 |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/README_pt.md b/README_pt.md
index 2b0b184c..2bc76a7b 100644
--- a/README_pt.md
+++ b/README_pt.md
@@ -46,8 +46,8 @@ Reconhecimento, sintese e compreensao de fala no dispositivo para Mac e iOS. Exe
 - **[VibeVoice TTS](https://soniqo.audio/pt/guides/vibevoice)** — TTS de formato longo / multi-alto-falante (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, sintese de podcast/audiolivro de ate 90 min, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/pt/guides/magpie)** — TTS multilíngue (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB ou CoreML INT8 342 MB, 9 idiomas, 5 oradores predefinidos, streaming em MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — TTS de flow-matching no dispositivo (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 idiomas, 10 vozes, sem G2P, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS multilíngue com clonagem de voz zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 idiomas, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — TTS de difusao nao-autorregressiva com clonagem de voz zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX int8 ~1 GB / fp16, 600+ idiomas, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS multilíngue com clonagem de voz zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 idiomas em runtime; hebraico requer niqqud, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — TTS de difusao nao-autorregressiva com clonagem de voz zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX fp16 por padrão / int8 disponível, 600+ idiomas, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLMs e tradução**
@@ -168,7 +168,8 @@ Vista compacta abaixo. **[Catalogo completo de modelos com tamanhos, quantizacoe
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/pt/guides/vibevoice) | Texto → Fala (formato longo, multi-alto-falante) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/pt/guides/vibevoice) | Texto → Fala (podcast de ate 90 min) | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/pt/guides/magpie) | Texto → Fala (5 oradores predefinidos, streaming) | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9 (CoreML exclui JA) |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | Texto → Fala (difusao NAR, clonagem zero-shot) | MLX | 0.8B (int8/fp16) | 600+ |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Texto → Fala (clonagem zero-shot) | MLX | 0.8B (fp16) | 23 (HE requer niqqud) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Texto → Fala (difusao NAR, clonagem zero-shot) | MLX | 0.8B (fp16 por padrão / int8) | 600+ |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Texto → Fala (clonagem zero-shot, marcadores de estilo explicitos) | MLX | 0.5B-class (fp16) | Multilingue |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/README_ru.md b/README_ru.md
index bfe0083e..4582a68a 100644
--- a/README_ru.md
+++ b/README_ru.md
@@ -46,8 +46,8 @@
 - **[VibeVoice TTS](https://soniqo.audio/ru/guides/vibevoice)** — Длинный формат / многоголосый TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, синтез подкастов/аудиокниг до 90 минут, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/ru/guides/magpie)** — Многоязычный TTS (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 МБ или CoreML INT8 342 МБ, 9 языков, 5 встроенных голосов, стриминг на MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — Синтез речи на устройстве на основе flow-matching (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 язык, 10 голосов, G2P-free, 44.1 кГц)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Многоязычный TTS с zero-shot клонированием голоса (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 ГБ, 23 языка, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — Неавторегрессионный диффузионный TTS с zero-shot клонированием голоса (k2-fsa OmniVoice, бэкбон Qwen3, MLX int8 ~1 ГБ / fp16, 600+ языков, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Многоязычный TTS с zero-shot клонированием голоса (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 ГБ, 23 языка в рантайме; для иврита нужен никкуд, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — Неавторегрессионный диффузионный TTS с zero-shot клонированием голоса (k2-fsa OmniVoice, бэкбон Qwen3, MLX fp16 по умолчанию / доступен int8, 600+ языков, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM и перевод**
@@ -168,7 +168,8 @@ struct DictateView: View {
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/ru/guides/vibevoice) | Текст → Речь (длинный формат, многоголосый) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/ru/guides/vibevoice) | Текст → Речь (подкаст до 90 минут) | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/ru/guides/magpie) | Текст → Речь (5 встроенных голосов, стриминг) | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9 (CoreML без JA) |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | Текст → Речь (NAR-диффузия, zero-shot клонирование) | MLX | 0.8B (int8/fp16) | **600+** |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Текст → Речь (zero-shot клонирование) | MLX | 0.8B (fp16) | 23 (HE нужен никкуд) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Текст → Речь (NAR-диффузия, zero-shot клонирование) | MLX | 0.8B (fp16 по умолчанию / int8) | **600+** |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Текст → Речь (zero-shot клонирование, явные маркеры стиля) | MLX | 0.5B-class (fp16) | Многоязычный |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/README_th.md b/README_th.md
index 2735185a..1a335ed9 100644
--- a/README_th.md
+++ b/README_th.md
@@ -46,8 +46,8 @@
 - **[VibeVoice TTS](https://soniqo.audio/guides/vibevoice)** — TTS แบบยาว / หลายผู้พูด (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX สังเคราะห์พอดแคสต์/หนังสือเสียงได้นานสูงสุด 90 นาที, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/guides/magpie)** — TTS หลายภาษา (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB หรือ CoreML INT8 342 MB, 9 ภาษา, 5 ผู้พูดสำเร็จรูป สตรีมมิ่งบน MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — TTS แบบ flow-matching บนอุปกรณ์ (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 ภาษา, 10 เสียง, G2P-free, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS หลายภาษาพร้อมการโคลนเสียงแบบ zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 23 ภาษา, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — TTS แบบ diffusion non-autoregressive พร้อมการโคลนเสียงแบบ zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX int8 ~1 GB / fp16, 600+ ภาษา, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS หลายภาษาพร้อมการโคลนเสียงแบบ zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1.3 GB, 23 ภาษาที่รันไทม์; ภาษาฮีบรูต้องมี niqqud, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — TTS แบบ diffusion non-autoregressive พร้อมการโคลนเสียงแบบ zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX fp16 เป็นค่าเริ่มต้น / มี int8, 600+ ภาษา, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM และการแปล**
@@ -168,7 +168,8 @@ struct DictateView: View {
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/guides/vibevoice) | ข้อความ → เสียงพูด (รูปแบบยาว หลายผู้พูด) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/guides/vibevoice) | ข้อความ → เสียงพูด (พอดแคสต์ยาวสุด 90 นาที) | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/guides/magpie) | ข้อความ → เสียงพูด (5 ผู้พูดสำเร็จรูป สตรีมมิ่ง) | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9 (CoreML ไม่รวม JA) |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | ข้อความ → เสียงพูด (diffusion NAR, การโคลนแบบ zero-shot) | MLX | 0.8B (int8/fp16) | 600+ |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | ข้อความ → เสียงพูด (โคลนแบบ zero-shot) | MLX | 0.8B (fp16) | 23 (HE ต้องมี niqqud) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | ข้อความ → เสียงพูด (diffusion NAR, การโคลนแบบ zero-shot) | MLX | 0.8B (fp16 ค่าเริ่มต้น / int8) | 600+ |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | ข้อความ → เสียงพูด (การโคลนแบบ zero-shot, marker สไตล์แบบชัดเจน) | MLX | 0.5B-class (fp16) | หลายภาษา |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/README_tr.md b/README_tr.md
index 7f1d88ef..49942861 100644
--- a/README_tr.md
+++ b/README_tr.md
@@ -46,8 +46,8 @@ Mac ve iOS için cihaz üzerinde konuşma tanıma, sentezleme ve anlama. Apple S
 - **[VibeVoice TTS](https://soniqo.audio/guides/vibevoice)** — Uzun biçimli / çok konuşmacılı TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, 90 dakikaya kadar podcast/sesli kitap sentezi, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/guides/magpie)** — Çok dilli TTS (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB veya CoreML INT8 342 MB, 9 dil, 5 hazır konuşmacı, MLX'te akış)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — Cihaz üzerinde flow-matching TTS (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 dil, 10 ses, G2P-free, 44.1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Zero-shot ses klonlama özellikli çok dilli TTS (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 dil, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — Zero-shot ses klonlama ile otoregresif olmayan difüzyon TTS (k2-fsa OmniVoice, Qwen3 omurgası, MLX int8 ~1 GB / fp16, 600+ dil, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — Zero-shot ses klonlama özellikli çok dilli TTS (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 çalışma zamanı dili; İbranice için niqqud gerekir, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — Zero-shot ses klonlama ile otoregresif olmayan difüzyon TTS (k2-fsa OmniVoice, Qwen3 omurgası, MLX fp16 varsayılan / int8 kullanılabilir, 600+ dil, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM ve çeviri**
@@ -168,7 +168,8 @@ Aşağıda kompakt bir görünüm. **[Boyutlar, kuantizasyonlar, indirme URL'ler
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/guides/vibevoice) | Metin → Konuşma (uzun biçimli, çok konuşmacılı) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/guides/vibevoice) | Metin → Konuşma (90 dakikaya kadar podcast) | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/guides/magpie) | Metin → Konuşma (5 hazır konuşmacı, akış) | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9 (CoreML, JA hariç) |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | Metin → Konuşma (NAR difüzyon, zero-shot klonlama) | MLX | 0.8B (int8/fp16) | **600+** |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Metin → Konuşma (zero-shot klonlama) | MLX | 0.8B (fp16) | 23 (HE için niqqud gerekir) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Metin → Konuşma (NAR difüzyon, zero-shot klonlama) | MLX | 0.8B (fp16 varsayılan / int8) | **600+** |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Metin → Konuşma (zero-shot klonlama, açık stil işaretleri) | MLX | 0.5B-class (fp16) | Çok dilli |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/README_vi.md b/README_vi.md
index 8a08f3e3..8e2555c6 100644
--- a/README_vi.md
+++ b/README_vi.md
@@ -46,8 +46,8 @@ Nhận dạng, tổng hợp và hiểu giọng nói trên thiết bị cho Mac v
 - **[VibeVoice TTS](https://soniqo.audio/guides/vibevoice)** — TTS định dạng dài / nhiều người nói (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, tổng hợp podcast/sách nói lên đến 90 phút, EN/ZH)
 - **[Magpie TTS](https://soniqo.audio/guides/magpie)** — TTS đa ngôn ngữ (NVIDIA Magpie-TTS Multilingual 357M, MLX INT8 411 MB hoặc CoreML INT8 342 MB, 9 ngôn ngữ, 5 giọng nói có sẵn, streaming trên MLX)
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — TTS flow-matching trên thiết bị (Supertone Supertonic-3 99M, CoreML/Neural Engine, 31 ngôn ngữ, 10 giọng, không cần G2P, 44,1 kHz)
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS đa ngôn ngữ với nhân bản giọng nói zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 ngôn ngữ, MIT)
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — TTS khuếch tán phi tự hồi quy với nhân bản giọng nói zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX int8 ~1 GB / fp16, 600+ ngôn ngữ, Apache-2.0)
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — TTS đa ngôn ngữ với nhân bản giọng nói zero-shot (Resemble AI Chatterbox Multilingual, MLX fp16 ~1,3 GB, 23 ngôn ngữ runtime; tiếng Hebrew cần niqqud, MIT)
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — TTS khuếch tán phi tự hồi quy với nhân bản giọng nói zero-shot (k2-fsa OmniVoice, backbone Qwen3, MLX fp16 mặc định / có int8, 600+ ngôn ngữ, Apache-2.0)
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM và dịch thuật**
@@ -168,7 +168,8 @@ Xem tổng quan gọn bên dưới. **[Danh mục mô hình đầy đủ với k
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/guides/vibevoice) | Văn bản → Giọng nói (định dạng dài, nhiều người nói) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/guides/vibevoice) | Văn bản → Giọng nói (podcast đến 90 phút) | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/guides/magpie) | Văn bản → Giọng nói (5 giọng có sẵn, streaming) | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9 (CoreML loại trừ JA) |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | Văn bản → Giọng nói (khuếch tán NAR, nhân bản zero-shot) | MLX | 0.8B (int8/fp16) | **600+** |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | Văn bản → Giọng nói (nhân bản zero-shot) | MLX | 0.8B (fp16) | 23 (HE cần niqqud) |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | Văn bản → Giọng nói (khuếch tán NAR, nhân bản zero-shot) | MLX | 0.8B (fp16 mặc định / int8) | **600+** |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | Văn bản → Giọng nói (nhân bản zero-shot, marker phong cách tường minh) | MLX | 0.5B-class (fp16) | Đa ngôn ngữ |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/README_zh.md b/README_zh.md
index 02053b5d..3f2ef3a8 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -46,8 +46,8 @@
 - **[VibeVoice TTS](https://soniqo.audio/zh/guides/vibevoice)** — 长篇 / 多说话人 TTS（Microsoft VibeVoice Realtime-0.5B + 1.5B，MLX，可合成最长 90 分钟的播客 / 有声书，英语 / 中文）
 - **[Magpie TTS](https://soniqo.audio/zh/guides/magpie)** — 多语言 TTS（NVIDIA Magpie-TTS Multilingual 357M，MLX INT8 411 MB 或 CoreML INT8 342 MB，9 种语言，5 位预设说话人，MLX 端流式）
 - **[Supertonic TTS](https://soniqo.audio/guides/supertonic)** — 端侧流匹配 TTS（Supertone Supertonic-3 99M，CoreML/神经引擎，31 种语言，10 种音色，G2P-free，44.1 kHz）
-- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — 支持零样本声音克隆的多语言 TTS（Resemble AI Chatterbox Multilingual，MLX fp16 ~1.3 GB，23 种语言，MIT）
-- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-int8)** — 支持零样本语音克隆的非自回归扩散 TTS（k2-fsa OmniVoice，Qwen3 骨干，MLX int8 ~1 GB / fp16，600+ 种语言，Apache-2.0）
+- **[Chatterbox TTS](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16)** — 支持零样本声音克隆的多语言 TTS（Resemble AI Chatterbox Multilingual，MLX fp16 ~1.3 GB，23 种运行时语言；希伯来语需要 niqqud，MIT）
+- **[OmniVoice TTS](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16)** — 支持零样本语音克隆的非自回归扩散 TTS（k2-fsa OmniVoice，Qwen3 骨干，MLX fp16 默认 / 可用 int8，600+ 种语言，Apache-2.0）
 - **[Indic-Mio](docs/models/indic-mio-tts.md)** — Hindi/Indic TTS with inline emotion markers and optional reference-voice cloning (MLX, 24 kHz)
 
 **LLM 与翻译**
@@ -168,7 +168,8 @@ struct DictateView: View {
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/zh/guides/vibevoice) | 文本 → 语音（长篇、多说话人） | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/zh/guides/vibevoice) | 文本 → 语音（最长 90 分钟播客） | MLX | 1.5B | EN/ZH |
 | [Magpie-TTS Multilingual](https://soniqo.audio/zh/guides/magpie) | 文本 → 语音（5 位预设说话人，流式） | MLX / CoreML | 357M (MLX INT8, CoreML INT8) | 9（CoreML 不含日语） |
-| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-int8) | 文本 → 语音（NAR 扩散，零样本克隆） | MLX | 0.8B (int8/fp16) | **600+** |
+| [Chatterbox Multilingual](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16) | 文本 → 语音（零样本克隆） | MLX | 0.8B (fp16) | 23（HE 需要 niqqud） |
+| [OmniVoice](https://huggingface.co/aufklarer/OmniVoice-MLX-fp16) | 文本 → 语音（NAR 扩散，零样本克隆） | MLX | 0.8B (fp16 默认 / int8) | **600+** |
 | [Indic-Mio](docs/models/indic-mio-tts.md) | Text → Speech (Hindi/Indic, emotion tags, voice cloning) | MLX | fp16 | Hindi / Indic |
 | [Fish Audio S2 Pro](docs/models/fish-audio-s2-pro.md) | 文本 → 语音（零样本克隆，显式风格标记） | MLX | 0.5B-class (fp16) | 多语言 |
 | [Qwen3.5 Chat](docs/models/qwen35-chat.md) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
diff --git a/Sources/ChatterboxTTS/ChatterboxTTSModel.swift b/Sources/ChatterboxTTS/ChatterboxTTSModel.swift
index 66508d6c..e82768e0 100644
--- a/Sources/ChatterboxTTS/ChatterboxTTSModel.swift
+++ b/Sources/ChatterboxTTS/ChatterboxTTSModel.swift
@@ -19,12 +19,15 @@ import MLXNN
 public enum ChatterboxModelError: Error, LocalizedError {
     case missingFile(String)
     case unsupportedLanguage(String)
+    case invalidText(String)
 
     public var errorDescription: String? {
         switch self {
         case let .missingFile(p): return "Chatterbox: required file not found: \(p)"
         case let .unsupportedLanguage(l):
             return "Chatterbox: language '\(l)' is not supported"
+        case let .invalidText(message):
+            return "Chatterbox: \(message)"
         }
     }
 }
@@ -333,7 +336,13 @@ public final class ChatterboxTTSModel {
         let t3PromptTokens = Array(s3gen.tokenizer.encode(ref16kEnc).prefix(Self.speechCondPromptLen))
 
         // --- text tokenization: [sot] + ids + [eot] ---
-        let ids = tokenizer.encode(text, languageId: lang)
+        let ids: [Int]
+        do {
+            ids = try tokenizer.encodeStrict(text, languageId: lang)
+        } catch ChatterboxTokenizerError.hebrewRequiresDiacritics {
+            throw ChatterboxModelError.invalidText(
+                "Hebrew input must include niqqud/diacritics; automatic Dicta diacritization is not bundled yet")
+        }
         let textTokens = [Self.startTextToken] + ids + [Self.stopTextToken]
 
         // --- T3: text -> speech tokens ---
diff --git a/Sources/ChatterboxTTS/MTLTokenizer.swift b/Sources/ChatterboxTTS/MTLTokenizer.swift
index 5162a3ab..40d3b22a 100644
--- a/Sources/ChatterboxTTS/MTLTokenizer.swift
+++ b/Sources/ChatterboxTTS/MTLTokenizer.swift
@@ -10,9 +10,21 @@ public enum ChatterboxToken {
     public static let space = "[SPACE]"
 }
 
-public enum ChatterboxTokenizerError: Error {
+public enum ChatterboxTokenizerError: Error, LocalizedError {
     case malformedTokenizerJSON
     case missingCangjieMap
+    case hebrewRequiresDiacritics
+
+    public var errorDescription: String? {
+        switch self {
+        case .malformedTokenizerJSON:
+            return "Malformed Chatterbox tokenizer.json"
+        case .missingCangjieMap:
+            return "Chatterbox tokenizer is missing Cangjie5_TC.json"
+        case .hebrewRequiresDiacritics:
+            return "Chatterbox Hebrew input must include niqqud/diacritics; automatic diacritization is not bundled yet"
+        }
+    }
 }
 
 /// Multilingual grapheme tokenizer for Chatterbox — a self-contained Swift port
@@ -50,11 +62,11 @@ public final class MTLTokenizer {
     ]
 
     /// Languages whose text frontend is implemented locally without falling
-    /// back to raw text. Keep languages with missing external frontends out of
-    /// this runtime gate: `he` needs Dicta-style diacritization.
+    /// back to raw text. Hebrew is accepted only when the caller supplies
+    /// pre-diacritized text; automatic Dicta ONNX diacritization is not ported.
     public static let supportedLanguages: Set<String> = [
-        "ar", "da", "de", "el", "en", "es", "fi", "fr", "hi", "it",
-        "ja", "ko", "ms", "nl", "no", "pl", "pt", "ru", "sv", "sw",
+        "ar", "da", "de", "el", "en", "es", "fi", "fr", "he", "hi",
+        "it", "ja", "ko", "ms", "nl", "no", "pl", "pt", "ru", "sv", "sw",
         "tr", "zh",
     ]
 
@@ -129,6 +141,19 @@ public final class MTLTokenizer {
         return tokenize(s)
     }
 
+    /// Strict encode path used by synthesis. It preserves `encode`'s legacy
+    /// behaviour for tests/debugging, but rejects Hebrew text that has not been
+    /// pre-diacritized; upstream relies on Dicta before tokenization for this.
+    public func encodeStrict(_ text: String, languageId: String? = nil) throws -> [Int] {
+        let lang = languageId?.trimmingCharacters(in: .whitespacesAndNewlines).lowercased()
+        if lang == "he",
+           Self.containsHebrewLetters(text),
+           !Self.containsHebrewDiacritics(text) {
+            throw ChatterboxTokenizerError.hebrewRequiresDiacritics
+        }
+        return encode(text, languageId: languageId)
+    }
+
     /// Inverse of `encode` for debugging.
     public func decode(_ ids: [Int]) -> String {
         var s = ids.map { idToToken[$0] ?? "" }.joined()
@@ -276,6 +301,27 @@ public final class MTLTokenizer {
         }
     }
 
+    public static func containsHebrewLetters(_ text: String) -> Bool {
+        text.unicodeScalars.contains { scalar in
+            (0x05D0...0x05EA).contains(scalar.value)
+        }
+    }
+
+    public static func containsHebrewDiacritics(_ text: String) -> Bool {
+        text.unicodeScalars.contains { scalar in
+            switch scalar.value {
+            case 0x0591...0x05BD, 0x05BF, 0x05C1...0x05C2, 0x05C4...0x05C5, 0x05C7:
+                return true
+            default:
+                return false
+            }
+        }
+    }
+
+    public static func isHebrewTextDiacritized(_ text: String) -> Bool {
+        !containsHebrewLetters(text) || containsHebrewDiacritics(text)
+    }
+
     private static func decomposeHangulToJamo(_ text: String) -> String {
         var out = ""
         for scalar in text.unicodeScalars {
diff --git a/Tests/ChatterboxTTSTests/MTLTokenizerTests.swift b/Tests/ChatterboxTTSTests/MTLTokenizerTests.swift
index 0aafaa02..6c3a797c 100644
--- a/Tests/ChatterboxTTSTests/MTLTokenizerTests.swift
+++ b/Tests/ChatterboxTTSTests/MTLTokenizerTests.swift
@@ -46,7 +46,9 @@ final class MTLTokenizerTests: XCTestCase {
         ]
         XCTAssertEqual(MTLTokenizer.upstreamLanguages, expected)
         XCTAssertEqual(MTLTokenizer.upstreamLanguages.count, 23)
-        XCTAssertFalse(MTLTokenizer.supportedLanguages.contains("he"))
+        XCTAssertEqual(MTLTokenizer.supportedLanguages, expected)
+        XCTAssertEqual(MTLTokenizer.supportedLanguages.count, 23)
+        XCTAssertTrue(MTLTokenizer.supportedLanguages.contains("he"))
         XCTAssertTrue(MTLTokenizer.supportedLanguages.contains("ja"))
         XCTAssertTrue(MTLTokenizer.supportedLanguages.contains("zh"))
 
@@ -105,6 +107,7 @@ final class MTLTokenizerTests: XCTestCase {
             "es": "Hoy probamos si el habla es clara y fácil de entender.",
             "fi": "Tänään testaamme, onko puhe selkeää ja helppoa ymmärtää.",
             "fr": "Aujourd'hui, nous testons si la parole est claire et facile à comprendre.",
+            "he": "הַיּוֹם אֲנַחְנוּ בּוֹדְקִים אִם הַדִּבּוּר בָּרוּר וְקַל לַהֲבָנָה.",
             "hi": "आज हम जाँचते हैं कि आवाज़ साफ़ और समझने में आसान है या नहीं।",
             "it": "Oggi testiamo se il parlato è chiaro e facile da capire.",
             "ja": "今日は音声が明瞭で理解しやすいかを確認します。",
@@ -123,8 +126,26 @@ final class MTLTokenizerTests: XCTestCase {
 
         XCTAssertEqual(Set(samples.keys), MTLTokenizer.supportedLanguages)
         for (language, text) in samples {
-            let encoded = tok.encode(text, languageId: language)
+            let encoded = try tok.encodeStrict(text, languageId: language)
             XCTAssertFalse(encoded.contains(1), "\(language) should not emit [UNK]")
         }
     }
+
+    func testHebrewRequiresDiacriticsForStrictEncoding() async throws {
+        let tok = try await loadTokenizer()
+
+        XCTAssertThrowsError(try tok.encodeStrict("שלום עולם", languageId: "he")) { error in
+            guard case ChatterboxTokenizerError.hebrewRequiresDiacritics = error else {
+                return XCTFail("expected Hebrew diacritics error, got \(error)")
+            }
+        }
+    }
+
+    func testHebrewDiacriticsEncodeWithoutUnknownTokens() async throws {
+        let tok = try await loadTokenizer()
+        let encoded = try tok.encodeStrict("שָׁלוֹם עוֹלָם", languageId: "he")
+
+        XCTAssertEqual(encoded.first, 2110)
+        XCTAssertFalse(encoded.contains(1), "diacritized Hebrew should not emit [UNK]")
+    }
 }
diff --git a/Tests/ChatterboxTTSTests/RoundtripE2ETests.swift b/Tests/ChatterboxTTSTests/RoundtripE2ETests.swift
index 4a0fd6d2..6b89df91 100644
--- a/Tests/ChatterboxTTSTests/RoundtripE2ETests.swift
+++ b/Tests/ChatterboxTTSTests/RoundtripE2ETests.swift
@@ -100,6 +100,7 @@ final class E2ERoundtripTests: XCTestCase {
         let cases: [(String, String)] = [
             ("en", "Hello there, this is a cloned voice."),
             ("ar", "مرحبا، هذا صوت مستنسخ."),
+            ("he", "שָׁלוֹם, זֶה קוֹל מְשֻׁבָּט."),
             ("hi", "नमस्ते, यह एक क्लोन की गई आवाज़ है।"),
             ("ja", "今日は音声が明瞭で理解しやすいかを確認します。"),
             ("zh", "今天我们测试中文语音是否清楚且容易理解。"),
diff --git a/docs/models/chatterbox-tts.md b/docs/models/chatterbox-tts.md
index dcac3e69..d0aa06bd 100644
--- a/docs/models/chatterbox-tts.md
+++ b/docs/models/chatterbox-tts.md
@@ -4,10 +4,10 @@
 
 Chatterbox Multilingual (Resemble AI, MIT) is a zero-shot voice-cloning TTS that
 synthesizes a target text in the timbre of a short reference clip. The upstream
-model publishes 23 language tags; the Swift runtime enables the 22 languages
-whose required text frontends are implemented locally. The MLX bundle is a
-genuine fp16
-conversion (~1.3 GB) of the three upstream checkpoints, published at
+model publishes 23 language tags; the Swift runtime enables all 23, with Hebrew
+requiring pre-diacritized text (niqqud) until automatic Dicta ONNX
+diacritization is bundled. The MLX bundle is a genuine fp16 conversion
+(~1.3 GB) of the three upstream checkpoints, published at
 [`aufklarer/Chatterbox-Multilingual-MLX-fp16`](https://huggingface.co/aufklarer/Chatterbox-Multilingual-MLX-fp16).
 
 The Swift port is built component-by-component, with token ids, mel features, and
@@ -16,7 +16,7 @@ speaker embeddings checked against a known-good reference at each stage.
 ## Languages
 
 Runtime-enabled languages: Arabic, Chinese, Danish, Dutch, English, Finnish,
-French, German, Greek, Hindi, Italian, Japanese, Korean, Malay, Norwegian,
+French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian,
 Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Turkish.
 
 Text encodes through a grapheme path: NFKD normalization → a per-language
@@ -25,8 +25,8 @@ Text encodes through a grapheme path: NFKD normalization → a per-language
 - Chinese uses word segmentation plus the published Cangjie-5 token map.
 - Japanese converts kanji readings to hiragana through `CFStringTokenizer`.
 - Korean decomposes Hangul syllables to Jamo.
-- Hebrew remains disabled until the runtime includes the diacritization frontend
-  needed by the upstream model.
+- Hebrew accepts pre-diacritized input with niqqud. Raw Hebrew is rejected with a
+  clear error until the runtime includes automatic Dicta ONNX diacritization.
 
 ## Pipeline Stages